4 ms·

I opened the link expecting a discussion of flawed metrics that fail to capture uncertainty or failure in replication of claimed studies, or complex architectur

I opened the link expecting a discussion of flawed metrics that fail to capture uncertainty or failure in replication of claimed studies, or complex architecture turning out to be shallow combinations of simpler ones. Instead the post attacks a straw man argument about "General intelligence", something that no AI researcher is claiming. Also the author is clearly clueless as evidenced by: "Ask a medical researcher to estimate the odds that a Google-funded AI-driven cancer moonshot will succeed, and I expect you will be met with a snicker. ' Google employs several researchers [1] who are well versed in both medicine as well as ML. Who are able to competently judge capabilities and limits of their tools and challenges involved in doing medical research. Rather than creating straw mental images of "medical researcher" "snickering" at "moonshots", the author should to engage in some self reflection before writing about scientific fields he has zero knowledge of. AND "I promise you, Facebook did not just work it all out in PHP, and they will not do so in my lifetime or yours." I won't comment on absurdity of the paragraph about FB or language & meaning. But if any reader was confused, Facebook AI team uses Torch/Lua (for deep learning) or Python (in FBLearnerFlow) and surely not PHP. [1] http://www.nimh.nih.gov/about/dr-tom-insel-to-step-down-as-nimh-director.shtml http://www.nimh.nih.gov/about/dr-tom-insel-to-step-down-as-n...

taneq 10y ago
If the odds aren't long enough to be met with a bit of a snicker, it's not a moonshot.
- harry8 10y ago
  Right a moonshot is spectacular, a triumph of PR, a source of wonder and ultimately a really stupid and unbelievably wasteful way to spend a vast fortune. Moonshot really should be the byword of how /not/ to do science funding. Likewise anything beginning "war on..." The war on cancer is going about as well as the war on drugs. Progress is made elsewhere. Moonshot odds weren't particularly long given the resourcing and desire and this makes it no less impressive as a feat of engineerning and resourcing. And no more sensible.
  - Bartweiss 10y ago
    I'm not convinced this is a general principle like "war on..." Moonshots are intrinsically unlikely, and I suppose we shouldn't expect lots of ancillary benefits (if you get those, it's "fundamental research" instead of a moonshot). But I don't think there's any requirement that moonshots be heavily financed, or otherwise carry net-negative expectation. They may not be the best social expenditure of funds, but those expenditures (typically attacking acute poverty) aren't realistically competing with moonshots for funding. Throwing a trickle of funding at long-odds projects seems like a worthwhile idea - I'll bet there are 100 different large-upside projects out there with a 1% chance of success.
    - taneq 10y ago
      Now that you mention it, I think the term 'moonshot' does imply the expectation of some kind of spin-offs being discovered along the way. After all, the original moon missions produced a whole range of useful inventions. I agree that it shouldn't be a pejorative term. The bigger a company is, the closer to basic research they can afford to go in order to create new products and markets. Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics.
      - Bartweiss 10y ago
        > Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics. I've always had the sense that they operate on a strategy of "if it's technological and important, we should be doing it". Self-driving cars and longevity enhancement aren't terribly natural markets for Google, but any successes will be worth a fortune. It seems like "don't miss anything" is the basic plan.
YeGoblynQueenne 10y ago
>> Instead the post attacks a straw man argument about "General intelligence", something that no AI researcher is claiming. Several of the Deep Learning people are on record making silly claims about the future of AI. I'll dig you up the links if you insist [1] but frex, Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that. Jeoff Hinton has said similar things. The reason for that is that the original AI project started out to create strong AI in the first place, so if you could not show that your theory has some sort of chance to lead to strong AI people would scoff at it. There's still a bit of that in academia and in industry and parts of the press have not yet gotten the memo that things have changed now, so you'll still hear lots of people making these claims, or having such claims drawn out of them in interviews. But that's besides the point- which is that a lot of the very clearly stated goals of the industry require the development of systems with human-like intelligence. For instance, natural language processing. We have made some advances in recent years (for instance in things like part of speech tagging or even machine translation to an extent) but the parts of language that we really take for granted as humans, conveying meaning and being aware of some sort of context, those are completely outside our grasp for the time being. So when Google pretends it can just throw a ton of processing power at, say, machine translation between any two languages any time, and get it to work well, you know that's just its marketing people knowing they can claim anything they like and only their engineering team will be in a position to know how far it is from the truth. _______ [1] please don't, I need to stop procrastinating.
- cLeEOGPw 10y ago
  > Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that. The only thing wrong in this sentence is your sarcastic tone. Neural nets, deep or not, rely on intuition to produce output, so it is completely true that deep nets show intuition. What is intuition? It's a feeling that something is of a certain way that is not obvious or immediately obvious/self evident. And translated to ML terms, it is some amount of certainty about something. Intuition is an intrinsic feature of any ML algorithm.
  - YeGoblynQueenne 10y ago
jeffmcmahan 10y ago
Here's the kind of claim I mean to attack in the polemic: "The algorithm's understanding of language 'has moved from a 2-year-old infant to something close to an 8 or 10-year-old child,' said Amit Singhal, a Google Fellow, an honorific reserved for the company's top engineers." Here's what shapes my thinking about medicine, moonshots, and Google: https://www.statnews.com/2016/06/06/google-star-trek-fiction/ https://www.statnews.com/2016/06/06/google-star-trek-fiction... About PHP - I am aware that no one would actually try to implement this kind of thing in PHP. I was going for snark. I probably shouldn't have.

taneq 10y ago

If the odds aren't long enough to be met with a bit of a snicker, it's not a moonshot.

harry8 10y ago
Right a moonshot is spectacular, a triumph of PR, a source of wonder and ultimately a really stupid and unbelievably wasteful way to spend a vast fortune. Moonshot really should be the byword of how /not/ to do science funding. Likewise anything beginning "war on..." The war on cancer is going about as well as the war on drugs. Progress is made elsewhere. Moonshot odds weren't particularly long given the resourcing and desire and this makes it no less impressive as a feat of engineerning and resourcing. And no more sensible.
- Bartweiss 10y ago
  I'm not convinced this is a general principle like "war on..." Moonshots are intrinsically unlikely, and I suppose we shouldn't expect lots of ancillary benefits (if you get those, it's "fundamental research" instead of a moonshot). But I don't think there's any requirement that moonshots be heavily financed, or otherwise carry net-negative expectation. They may not be the best social expenditure of funds, but those expenditures (typically attacking acute poverty) aren't realistically competing with moonshots for funding. Throwing a trickle of funding at long-odds projects seems like a worthwhile idea - I'll bet there are 100 different large-upside projects out there with a 1% chance of success.
  - taneq 10y ago
    Now that you mention it, I think the term 'moonshot' does imply the expectation of some kind of spin-offs being discovered along the way. After all, the original moon missions produced a whole range of useful inventions. I agree that it shouldn't be a pejorative term. The bigger a company is, the closer to basic research they can afford to go in order to create new products and markets. Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics.
    - Bartweiss 10y ago
      > Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics. I've always had the sense that they operate on a strategy of "if it's technological and important, we should be doing it". Self-driving cars and longevity enhancement aren't terribly natural markets for Google, but any successes will be worth a fortune. It seems like "don't miss anything" is the basic plan.

YeGoblynQueenne 10y ago

>> Instead the post attacks a straw man argument about "General intelligence", something that no AI researcher is claiming. Several of the Deep Learning people are on record making silly claims about the future of AI. I'll dig you up the links if you insist [1] but frex, Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that. Jeoff Hinton has said similar things. The reason for that is that the original AI project started out to create strong AI in the first place, so if you could not show that your theory has some sort of chance to lead to strong AI people would scoff at it. There's still a bit of that in academia and in industry and parts of the press have not yet gotten the memo that things have changed now, so you'll still hear lots of people making these claims, or having such claims drawn out of them in interviews. But that's besides the point- which is that a lot of the very clearly stated goals of the industry require the development of systems with human-like intelligence. For instance, natural language processing. We have made some advances in recent years (for instance in things like part of speech tagging or even machine translation to an extent) but the parts of language that we really take for granted as humans, conveying meaning and being aware of some sort of context, those are completely outside our grasp for the time being. So when Google pretends it can just throw a ton of processing power at, say, machine translation between any two languages any time, and get it to work well, you know that's just its marketing people knowing they can claim anything they like and only their engineering team will be in a position to know how far it is from the truth. _______ [1] please don't, I need to stop procrastinating.

cLeEOGPw 10y ago
> Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that. The only thing wrong in this sentence is your sarcastic tone. Neural nets, deep or not, rely on intuition to produce output, so it is completely true that deep nets show intuition. What is intuition? It's a feeling that something is of a certain way that is not obvious or immediately obvious/self evident. And translated to ML terms, it is some amount of certainty about something. Intuition is an intrinsic feature of any ML algorithm.
- YeGoblynQueenne 10y ago

jeffmcmahan 10y ago

Here's the kind of claim I mean to attack in the polemic: "The algorithm's understanding of language 'has moved from a 2-year-old infant to something close to an 8 or 10-year-old child,' said Amit Singhal, a Google Fellow, an honorific reserved for the company's top engineers." Here's what shapes my thinking about medicine, moonshots, and Google: https://www.statnews.com/2016/06/06/google-star-trek-fiction/ https://www.statnews.com/2016/06/06/google-star-trek-fiction... About PHP - I am aware that no one would actually try to implement this kind of thing in PHP. I was going for snark. I probably shouldn't have.